Cache Performance Analysis of Virtualized Router on Virtual Content-Centric Networks

International Journal of Next-Generation Networks (IJNGN) Vol.5, No.4, December 2013
DOI : 10.5121/ijngn.2013.5401 1
CACHE PERFORMANCE ANALYSIS OF
VIRTUALIZED ROUTER ON
VIRTUAL CONTENT-CENTRIC NETWORKS
Keiichiro Tsukamoto1
, Kaito Ohsugi1
, Hiroyuki Ohsaki2
,
Toru Hasegawa1
and Masayuki Murata 1
1
Graduate School of Information Science and Technology,
Osaka University, Osaka, Japan
2
Department of Informatics, School of Science and Technology,
Kwansei Gakuin University, Hyogo, Japan
ABSTRACT
Content-centric networking (CCN) is one of the major proposals for realizing information-centric
networking. CCN routers cache forwarded data in a buffer memory called the ContentStore (CS). Virtual
content-centric networking (VCCN), which enables the construction of multiple virtual networks (called
VCCN slices) on a content-centric network, has been recently proposed. When multiple VCCN slices are
constructed, the performance of each VCCN slice and that of the entire network are strongly affected by the
CCN routers' CS allocation to VCCN router instances in VCCN slices. In this paper, we analyze the effects
of CS allocation methods and content request patterns in VCCN slices on the performance of each VCCN
slice and that of the entire network. Through several numerical examples, we show that when content
request patterns are heterogeneous, a hybrid resource allocation method is effective in terms of both
network fairness for VCCN slices and overall network performance.
KEYWORDS
CCN (Content-Centric Networking), Router Virtualization, Resource Allocation, Markov Chain, LRU
(Least Recently Used)
1. INTRODUCTION
Content-centric networking (CCN), in which a router routes a packet based on an identifier as-
signed to content, is one of the major proposals for realizing information-centric networking [1].
In CCN, a request packet from a user, called an Interest packet, is routed between CCN routers
according to the longest prefix matching the requesting content identifier. If the Interest packet is
successfully delivered to the source, the content packet, called a Data packet, is sent back to the
user by traversing the path of the Interest packet. CCN routers cache forwarded Data packets in a
buffer memory called the ContentStore (CS). When a CCN router receives an Interest packet for a
cached Data packet, it returns the cached packet so that the average content delivery time and the
amount of traffic over the network can be reduced.
Virtual content-centric networking (VCCN), which enables the construction of multiple virtual
networks (called VCCN slices) on a content-centric network, has been recently proposed [2].

2
VCCN slices are constructed by operating a CCN router as multiple, logically independent
VCCN router instances and by logically connecting VCCN router instances that are not adjacent
in the network.
When multiple VCCN slices are constructed, the performance of each VCCN slice and that of the
entire network are strongly affected by the CCN routers’ resource allocation to VCCN router
instances in VCCN slices. Several previous studies have shown clearly that, in CCN, the
effectiveness of content caching depends strongly on the content request pattern experienced by
the CS of a CCN router [3–5]. Hence, the performance of each VCCN slice and that of the entire
network depend strongly on how a CCN router allocates its CS to VCCN router instances on
VCCN slices that have different content request patterns.
In this paper, we analytically investigate the effect of CS allocation methods and content request
patterns in VCCN slices in terms of the network fairness for VCCN slices and overall network
performance. We focus on the effects of the content popularity slope and the content request ratio
of each VCCN slice, which are the main features of a content request pattern and which
significantly affect the effectiveness of content caching in particular.
In this paper, we focus on three types of CS allocation methods: an exclusive method, a shared
method and a hybrid method. In the exclusive method, each VCCN router instance within a CCN
router monopolizes a given part of its CS. In the shared method, all VCCN router instances within
a CCN router use its entire CS jointly. In the hybrid method, several VCCN router instances
within a CCN router are assigned their own parts of its CS and other instances jointly use the
remaining CS. Previous studies of the effects of content caching on content-centric networks have
focused only on the exclusive and shared methods [6–9]. However, when content request patterns
are heterogeneous, these two methods can barely maintain a balance between network fairness for
VCCN slices and overall network performance. Hence, we conjectured that a hybrid method,
which has the characteristics of both the exclusive and shared approaches, might be a useful CS
allocation method on a content-centric network in which there are multiple content request
patterns in VCCN slices. In this paper, we quantitatively compare a hybrid method with the two
existing methods in terms of the fairness for VCCN slices and the overall network performance.
The main contribution of this paper is twofold. First, we develop a mathematical model of
virtualized CCN router for cache performance analysis under arbitrary content request patterns,
and derive the cache hit rate for each VCCN router instance and the aggregated cache hit rate of
the virtualized CCN router. Second, through numerical examples, we quantitatively show that in
di- verse scenarios, the hybrid method can provide desirable trade-offs among the network
fairness for VCCN slices and overall network performance.
The organization of this paper is as follows. Section 2 contains a summary of related work. In
Section 3, we describe CCN router virtualization and CS allocation to VCCN router instances. In
Section 4, our model of a virtualized CCN router which accommodates multiple VCCN router
instances is described and analytical results are derived. In Section 5, through several numerical
examples, we analyze the effects of CS allocation methods and content request patterns in VCCN
slices on the network fairness for VCCN slices and the overall network performance. Finally, in
Section 6, we give our conclusions and indicate the direction of future work.
2. RELATED WORK
The effect of content caching on content-centric networks where multiple applications or services
are running has been investigated in [6–9]. Carofiglio et al. [6, 7] have clarified the role of the CS
allocation method (an exclusive method) on the cache hit rate of multiple applications running on

3
content-centric networks by means of experiments and simulations. Their results show that an
exclusive method can guarantee application performance but it may decrease the overall
performance of the entire network. Fricker et al. [8] have evaluated the cache hit rate of multiple
services running concurrently on a content-centric network using an approximation proposed by
Che et al. [10]. Their results show that allowing a service with a rapid content popularity slope to
monopolize the CS raises the cache hit ratio rather than increasing the quantity of CS shared
between all the services, when the size of the CS is large. Ohsugi et al. [9] investigated by means
of a simulation the effect of CS allocation in CCN router virtualization (an exclusive method) in
terms of the average content delivery time for an entire content-centric network on which
multiple applications are running. Their results show that exclusive CS allocation in CCN router
virtualization increases the average content delivery time by about 20% in the worst case and
improves network fairness for applications.
These studies show that while a shared method is preferable for maximizing the performance of
the entire network, an exclusive method is preferable for improving network fairness for
applications or services. However, a hybrid method, which will strike a balance between
performance and fairness, has not yet been described or quantitatively evaluated.
The effect of a content request pattern in a content-centric network on caching performance has
been investigated in [3, 4]. Rossini et al. [3, 4] evaluated by means of simulations the dependence
of the cache hit rate on several aspects of network design, such as topology, content size, content
popularity, the locality of user requests and the number of repositories. Their results show that the
Zipf exponent α, representing content popularity, can have a dramatic impact on the performance
of the entire network. However, in those studies, CCN router virtualization was not taken into
consideration, and the relation between CS allocation methods and content request patterns in
VCCN slices has not yet been clarified.
3. CCN ROUTER VIRTUALIZATION AND CS ALLOCATION TO VCCN ROUTER
INSTANCES
With CCN router virtualization, a single CCN router is operated as multiple, logically
independent VCCN router instances (see Fig. 1 [2]). CCN router virtualization can be easily
realized by switching three data structures used for packet routing in CCN: the forwarding
information base (FIB), the CS, and the pending interest table (PIT) [2]. A VCCN router instance
routes packets using its own data structures (i.e., FIB, CS and PIT) in the same way that a CCN
router does.
A VCCN slice can be built by logically connecting VCCN router instances (see Fig. 1). In a
VCCN slice, packet transport between VCCN router instances, which are commonly not adjacent
in the CCN network, is enabled by lower layer protocols than those defining CCN or by tunneling
in the CCN layer [2].
In the following, we describe three types of conceivable methods of allocating CS resources to
VCCN router instances: an exclusive method, a shared method and a hybrid method (see Fig. 2).
• Exclusive method
Each VCCN router instance monopolizes a given part of the CS of a CCN router. One
advantage of the exclusive method is that the performance of content caching in a VCCN
router instance is independent of that in other VCCN router instances. On the other hand, a
disadvantage is that the cache miss rate may increase because a VCCN router instance

4
monopolizes a given part of the CS regardless of the amount of traffic in the VCCN router
instance.
• Shared method
All VCCN router instances jointly use the entire CS of a CCN router. One advantage of the
shared method is that there is no loss of CS due to splitting of the CS and the cache hit rate may
be increased because all VCCN router instances use the entire CS. On the other hand, a
disadvantage is that a VCCN router instance with a large amount of traffic may monopolize
most of the CS because each VCCN router instance affects the others within a given CCN
router.
C
F
G
D
B
E
A
VCCN router instance
for VCCN slice X
for VCCN slice Y
VCCN slice X
VCCN slice Y
link
CCN router
A
X
E
Y
B
Y
D
Y
E
X
Figure 1: An example of VCCN networks built on a CCN network: two logically independent VCCN
networks, X and Y, are built on a network composed of seven CCN routers, A through G.
R
2
R
1
R
3
1
2
3
4
5
...
C
Hybrid method
R
2
R
1
R
3
1
2
3
4
5
...
C
CCN router
R
2
R
1
R
3
1
2
3
4
5
...
C
CS of CCN router
Shared method
Exclusive method
Figure 2: Examples of methods for the allocation of CS to VCCN routers (an exclusive method, a shared
method and a hybrid method).
• Hybrid method

5
Several VCCN router instances within a CCN router are assigned their own parts of its CS and
other instances jointly use the remaining CS. Advantages of the hybrid method are that CS loss
is reduced relative to the exclusive method and each VCCN router instance has a minimal
effect on the others within a given CCN router. On the other hand, a disadvantage is that the
management of the CS may be more complicated than in either the exclusive method or the
shared method.
Another approach, in which each VCCN router instance monopolizes a small part of the CS and
all the instances jointly use the remaining CS, is also conceivable as a hybrid method. However,
in this approach, each VCCN router instance needs to be assigned a certain amount of CS. As the
number of VCCN router instances running on a CCN router increases, the required size of CS
also increases. Hence, this hybrid approach is not considered in this paper.
4. VIRTUALIZED CCN ROUTER MODEL
4.1. Model description and notation
Our virtualized CCN router model is presented in Fig. 3. VCCN slices Sn
(1 ≤ n ≤ N) are
constructed on a content-centric network, and VCCN router instances Rn
(1 ≤ n ≤ N)
corresponding to the VCCN slices operate on a virtualized CCN router.
We assume that the request arrival process for content c on VCCN router instances Rn
is Poisson
with mean arrival rate λn
c. Thus, neither the content popularity slope nor the content request ratio
for each VCCN slice vary dynamically. G denotes the number of content items requested from all
the VCCN router instances.
The CS of the virtualized CCN router is partitioned into M segments and segment m (1 ≤ m ≤ M)
of size C(m) is jointly used by all VCCN router instances belonging to a set Ω(m). The sets Ω(m)
satisfy the following relations.
So a VCCN router instance does not use multiple segments. Each segment employs a least
recently used (LRU) replacement policy.
In our model, each Interest and Data packet has size L in order to simplify the problem. If Data
packets have different sizes, we can use the methods of Fricker et al. [11].
Moreover, we do not consider the aggregation of requests for the same content on the virtualized
CCN router because request aggregation has no or little impact on the stationary average content
delivery time [12]. When a CCN router receives an Interest packet for content that is already
being requested, the CCN router prevents the dispatch of that Interest packet.
In addition, we assume that the processing times for managing the CS, writing Data packets into
the CS and reading Data packets from the CS are negligible.

6
4.2. Determination of the cache hit rate based on a Markov chain model
First, we derive the cache hit rate pn
for each VCCN router instance Rn
(1 ≤ n ≤ N) and the
aggregated cache hit rate p of the virtualized CCN router based on a Markov chain model.
Although it is difficult to derive the performance of a large-scale CCN network in this model
because of the huge computational complexity, the state distribution for each successive storage
of content in the CS is acquired. Hence, we can estimate the size of CS required to hold a specific
amount of content.
Here, we focus VCCN router instances Rn
∈ Ω(m) which jointly use segment m of the CS. We
denote the state, in which content c on VCCN slice Sn
is in the kth segment of the CS, by sc,k and
we consider the Markov chain composed of sc,k (0 ≤ k ≤ C(m)) (see Fig. 4). We assume without
loss of generality that the new content is written into the top of the CS segment. Note that the
state in which content c on VCCN slice Sn
is not in the CS segment is denoted by sc,0 and
transitions to the same state are omitted in Fig. 4.
Let Pc,k and Pc,C(m) be the transition rate from sc,k to sc,k+1 and from sc,C(m) to sc,0, respectively.
CCN router
R
2
R
1
R
N
...
...
1
2
3
4
5
...
C
CS of CCN router
Mean arrival rate of content
...
...
...
...
...
l1
= {l1
1, l1
2, ..., l1
G}
From
downstream
To
upstream
Physical link
Virtual link
C(1)
C(2)
C(M)
...
l2
= {l2
1, l2
2, ..., l2
G}
...
lN
= {lN
1 , lN
2 , ..., lN
G}
Figure 3: The model considered in this paper.
sc,1
sc,2
sc,k
sc,C(m)
sc,0
...
Lm
c
...
Pc,1 Pc,2
Pc,k
Pc,C(m)-1
Pc,C(m)
Lm
c
Lm
c
Lm
c
Pc,k-1
Figure 4: Markov chain model. The state in which content c on VCCN slice Sn
is placed in the kth segment
of the CS is denoted by sc,k (0 ≤ k ≤ C(m)).

7
Pc,k is given by
where Λm
c is the request arrival rate of content c at CS segment m which is given by
where Θk
c is the set composed of request rates for content in segment k, which satisfies Θk
c ⊂
and |Θk
c| = k.
Let πc,k (0 ≤ k ≤ C(m)) be the equilibrium probability of sc,k. Then πc,k is given by
The cache hit rate pn
c of content c on VCCN slice Sn
can be derived from πc,k.
Finally, the cache hit rate pn
(1 ≤ n ≤ N) and the aggregated
cache hit rate p of the virtualized CCN router are given by the following equations.
4.2. Determination of the hit rate using an approximation method
We can derive an approximation to the cache hit rate pn
(1 ≤ n
≤ N) and the aggregated cache hit rate p of the virtualized CCN router from the hierarchical Web
caching model [10]. In this approximation, the network performance of a large-scale CCN
network can also be derived.

8
If the size of the cache memory is C, the cache replacement policy is LRU and the request arrival
process for content c is Poisson with mean arrival rate λc, the cache hit rate of content c is given
by
where tc is called the characteristic time of content c and is defined as the maximum inter-arrival
time between two adjacent requests for content c without a cache miss at the cache [10]. The
characteristic time can be calculated by solving
where Fc(t < tc) is the cumulative distribution (1 − e−λc tc ) of the inter-arrival time for requests
for content c at the cache. Without loss of generality, suppose that the caching of content c occurs
at t = 0. Thus, the characteristic time tc is the time at the CS, whose size is C, will be filled with
content other than c. In addition, (10) and (11) can be simplified as follows [11].
where tC is found by solving
In our model (see Fig. 4.1), the CS of a virtualized CCN router is partitioned into M segments and
each segment runs independently. Hence, the cache hit rate for each item of content can be
obtained by applying the approximation [10] to each segment.
Thus, the cache hit rate pn
c of content c on VCCN slice Sn
is given by
where tm
c can be calculated by solving
These equations can also be simplified to
where tm
is found by solving

9
From the above, the cache hit rate pn
of each VCCN router instance Rn
(1 ≤ n ≤ N) and the
aggregated cache hit rate p of the virtualized CCN router are obtained:
If (12) and (13) are used, tm
c is replaced with tm
.
5. NUMERICAL EXAMPLE
5.1. Validation of the model
First, we validated our model by comparing the analytic results of our model with the simulation
results. In the determination of the cache hit rate using the approximation method, we used (16)
and (17). The network tested is shown in Fig. 5. Three (N = 3) VCCN slices S1
, S2
and S3
are
constructed on a content-centric network so that we can study three CS allocation methods (i.e.,
an exclusive method, a shared method and a hybrid method). The link delay between nodes is
10[ms] irrespective for all slices. In each slice, 10,000 content items (L = 1 [Mbyte]) are stored in
the repository (i.e., G = 30, 000). Users generate content requests for each VCCN router instance
according to a Poisson process of intensity λ = 5[req/s]. The distribution of content popularity is
Zipf with parameter αn for VCCN slice Sn
. The content request ratios in all VCCN slices are equal.
Here, α1, α2 and α3 are set for 0.5, 0.75 and 1.5, respectively, by referring to the values of α on
existing services [3, 8]. Content requests in VCCN slices do not overlap; this is similar to the
situation considered in [6, 8]. We consider five CS allocation methods: an exclusive method that
assigns an equal number of CS segments to each VCCN router instance, a shared method that
assigns the entire CS to all VCCN router instances and hybrid methods (Hybrid(Sn
(1 ≤ n ≤ 3))
that assign one-third of the CS segments to VCCN router instance Rn
and assign the remaining
segments to the others.
Figure 5: Network used for the evaluation.

10
The cache hit rates of VCCN router instance R1
and the aggregated cache hit rates of the
virtualized CCN router against the size of the CS are shown in Figs. 6 and 7. Fig. 6 shows the
results based on the Markov chain model and the simulation results for the exclusive method, the
shared method and the hybrid method. In Fig. 6, for convenience of computational complexity, G
was set 100. We confirmed that the tendency of content caching is not very different by this
reduction of the number of contents. Fig. 7 shows the results from the approximate analysis and
the simulation results for the exclusive method, the shared method and the hybrid method. Since
the 95% confidence intervals of simulation results are too small (all of them are less than 10-3
),
they are omitted in this paper.
The differences between the analysis results and the simulation results are small in terms of both
the VCCN slice’s performance and the overall network performance, with a maximum error less
than 2%. The results for the other VCCN router instances are similar but are omitted to save
space. Moreover, it can be seen from Figs. 6 and 7 that the approximate analysis is as accurate as
the Markov chain based analysis.
In all the following results, we convert the cache hit rate into the average content delivery time so
that it is easier to understand the direct impact on users. We define the average content delivery
time between the users and the virtualized CCN router as τ1 and that between the virtualized CCN
router and the repositories as τ2. The average content delivery times τ1 and τ2 include both a
transmission delay and a processing delay. The average content delivery time D of the entire
network is given by
Repositories
Users
R
3
3000
CS of CCN router (C = 3,000 Mbyte)
Link delay = 10 ms
Figure5: Network used for theevaluation.
Thecache hit rates of VCCN router instance R1
and theaggregated cache hit rates of thevirtual-
ized CCN router against the size of the CS are shown in Figs. 6 and 7. Fig. 6 shows the results
based on theMarkov chain model and the simulation results for theexclusivemethod, theshared
method and the hybrid method. In Fig. 6, for convenience of computational complexity, G was
set 100. Weconfirmed that thetendency of content caching isnot very different by thisreduction
of the number of contents. Fig. 7 shows the results from the approximate analysis and the simu-
lation results for theexclusivemethod, the shared method and the hybrid method. Since the95%
confidence intervals of simulation results are too small (all of them are less than 10− 3
), they are
omitted in thispaper.
The differences between theanalysis results and the simulation results are small in terms of both
the VCCN slice’s performance and the overall network performance, with a maximum error less
than 2%. Theresultsfor theother VCCN router instancesaresimilar but areomitted tosavespace.
Moreover, it can be seen from Figs. 6 and 7 that the approximate analysis is as accurate as the
Markov chain based analysis.
In all thefollowing results, weconvert thecachehit rateinto theaveragecontent delivery timeso
that it is easier to understand the direct impact on users. We define the average content delivery
timebetween theusersand thevirtualized CCN router asτ1 and that between thevirtualized CCN
router and the repositories as τ2. The average content delivery times τ1 and τ2 include both a
transmission delay and a processing delay. The average content delivery time D of the entire
network isgiven by
D =
N
n= 1
G
c= 1 2λn
c (τ1 + τ2(1 − pn
c ))
N
n= 1
G
c= 1 λn
c
(20)
Theanalytic solutionsfor theaveragecontent delivery timeD of theentirenetwork arealso highly
accurate, with amaximum error lessthan 2%.
5.2. Effectsof content popularity slopesin VCCN slices
Second, we investigated the effects of content popularity slopes in VCCN slices on fairness and
overall network performance. Figure 8 shows the average content delivery time of the entire
network and the fairness index [13] for VCCN slices against the difference of Zipf exponents α
between VCCN slices. Specifically, the figure shows the result of approximate analysis when
α1 = 1 − d, α2 = 1 and α3 = 1 + d for (0 ≤ d ≤ 1).
The analytic solutions for the average content delivery time D of the entire network are also
highly accurate, with a maximum error less than 2%.
5.2. Effects of content popularity slopes in VCCN slices
Second, we investigated the effects of content popularity slopes in VCCN slices on fairness and
overall network performance. Figure 8 shows the average content delivery time of the entire
network and the fairness index [13] for VCCN slices against the difference of Zipf exponents α
between VCCN slices. Specifically, the figure shows the result of approximate analysis when α1 =
1 − d, α2 = 1 and α3 = 1 + d for (0 ≤ d ≤ 1).
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16
Cache
hit
rate
ContentStore size [Mbyte]
VCCN router R
1
- model
VCCN router R
1
- simulation
CCN router - model
CCN router - simulation
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16
Cache
hit
rate
VCCN router R
1
- model
VCCN router R
1
- simulation
CCN router - model
(a) Exclusive (b) Shared

11
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16
Cache
hit
rate
VCCN router R
1
- model
VCCN router R
1
- simulation
CCN router - model
(c) Hybrid(S2
)
Figure 6: The cache hit rate against the size of CS for the exclusive method, the shared method and a hybrid
method (Markov chain based analysis).
Figure 8 shows that the overall network performance hardly depends on CS allocation methods
and that the shared method is preferable for improving fairness, on a content-centric network in
which there are slices with different content popularity slopes. As the difference between the
content popularity slopes increases, the size of the CS that S3
(with high α) requires to achieve a
high cache hit rate decreases and the size of the CS that S1
(with low α) requires increases. Hence,
as the difference between the content popularity slopes increases, an exclusive method cannot
keep up with the change of CS size required for each VCCN router instance and the fairness
index is degraded. At the least, this result indicates that when the difference of α between any
slices is 0.4 or more, the simple exclusive method is not suitable in terms of both fairness and
overall network performance.
5.3. Effects of content popularity slopes in VCCN slices
Third, we investigated the effects of content request ratios in VCCN slices on fairness and overall
network performance. Figure 9 shows the average content delivery time for the entire network
and the fairness index against the ratio r (0 ≤ r ≤ 1) of content requests for VCCN slice S1
compared with all requests. Specifically, the figure shows the results of the approximate analysis
when the content request ratios of VCCN slices S1
, S2
, S3
are respectively r, (1 − r)/2 and (1 − r)/2,
and α = 0.75 for all the VCCN slices.
Figure 9 shows that, on a content-centric network in which slices have very different content
request ratios, while a shared method is preferable for maximizing the performance of the entire
network, an exclusive method is preferable for improving fairness. This result is consistent with
existing studies [6,7,9]. When the content request ratio of VCCN slice S1
is higher than that of the
other slices (i.e., r ≥ 0.4), as r increases, the average content delivery time for the entire
0
0.2
0.4
0.6
0.8
1
0 1000 2000 3000 4000 5000 6000
Cache
hit
rate
VCCN router R
1
- model
VCCN router R
1
- simulation
CCN router - model
0
0.2
0.4
0.6
0.8
1
0 1000 2000 3000 4000 5000 6000
Cache
hit
rate
VCCN router R
1
- model
VCCN router R
1
- simulation
CCN router - model
(a) Exclusive (b) Shared

12
0
0.2
0.4
0.6
0.8
1
0 1000 2000 3000 4000 5000 6000
Cache
hit
rate
VCCN router R
1
- model
VCCN router R
1
- simulation
CCN router - model
(c) Hybrid(S2
)
Figure 7: The cache hit rate against the size of CS for the exclusive method, the shared method and a hybrid
method (approximate analysis).
network decreases for the shared method and the fairness index increases for the exclusive
method. On the other hand, Fig. 9 also shows that the hybrid methods (Hybrid(S2
) and
Hybrid(S3
)), which assign a part of the CS to slices other than S1
, achieve moderate performance
in terms of both metrics. At the least, this result indicates that when the content request ratio of a
specific slice is higher than that of the other slices, a hybrid method which assigns a part of the
CS to a slice with low content request ratio, is preferable for providing a balance between the two
metrics. From the above results we conclude that, when content request ratios in VCCN slices are
different, the hybrid method is best suited for providing a balance between fairness and overall
network performance.
5.4. Effects of content request patterns in VCCN slices
Finally, we investigated the combined effects of content popularity slope and content request ratio
in each VCCN slice on fairness and overall network performance. Figure 10 shows the average
content delivery time for the entire network and the fairness index against content request ratio r
(0 ≤ r ≤ 1) for VCCN slice S1
. The figure shows the results of the approximate analysis when α1,
α2 and α3 are 0.5, 0.75 and 1.5, respectively.
Figure 10 shows that, on a content-centric network in which there are slices with widely different
content popularity slopes and content request ratios the shared method is preferable for
maximizing the performance of the entire network, while a hybrid method (Hybrid(S2
)) is
preferable for improving fairness. As the content request ratio of S1
(which has low α) increases,
the size of the CS required by R1
is much larger than when only the content request ratio of S1
differs. Hence, the fairness index for the shared method is degraded due to the occupancy of the
CS by R1
, and that for the exclusive method is degraded for the reason given in Section 5.2. In
20
25
30
35
40
0 0.2 0.4 0.6 0.8 1
Average
content
delivery
time
[ms]
Difference in a between VCCN slices
Hybrid(S
1
)
Hybrid(S2
)
Hybrid(S
3
)
Exclusive
Shared
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
Fairness
index
Difference in a between VCCN slices
Hybrid(S
1
)
Hybrid(S
2
)
Hybrid(S3
)
Exclusive
Shared
(a) Average content delivery time of the entire network (b) Fairness index for VCCN slices
Figure 8: Average content delivery time and fairness index against the difference of Zipf exponent α
between VCCN slices, for each allocation method.

13
20
25
30
35
40
0 0.2 0.4 0.6 0.8 1
Average
content
delivery
time
[ms]
Content request ratio r of VCCN slice S
1
Hybrid(S1
)
Hybrid(S
2
)
Hybrid(S
3
)
Exclusive
Shared
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
Fairness
index
1
Hybrid(S1
)
Hybrid(S
2
)
Hybrid(S
3
)
Exclusive
Shared
Figure 9: Average content delivery time and fairness index against the ratio r of content requests for VCCN
slice S1
compared to all requests for each allocation method.
this case, in order to improve fairness, R1
and R3
, which require a small CS to achieve a high
cache hit rate, should jointly use a part of CS, and R2
should monopolize the remaining CS. At the
least, this result indicates that when the difference of α between any slices is approximately 1.0
and the content request ratio of a slice with low α is higher than those of the other slices (i.e., r ≥
0.4), the hybrid method which assigns a part of CS to a slice with a low content popularity slope
and low content request ratio, is preferable for improving fairness. From the above results, when
both content popularity slopes and content request ratios in VCCN slices are different, the hybrid
method is best suited for providing a balance between fairness and overall network performance.
For the hybrid methods, we also investigated the effect of the size of the CS assigned to a
VCCN slice on fairness and overall network performance. We used three CS allocation
methods: Hybrid(Sn
(1 ≤ n ≤ 3)) that assign a CS segment whose size is V to VCCN router
instance Rn
and assign the remaining CS to the others. Figure 11 shows the average content
delivery time for the entire network and the fairness index against V. Specifically, the figure
shows the results of the approximate analysis when r, α1, α2 and α3 are 0.9, 0.5, 0.75 and 1.5,
respectively.
Figure 11 shows that several hybrid methods perform well over a wide range of V. If V < 2000 on
Hybrid(S2
) or 1000 < V on Hybrid(S1
), the hybrid methods achieve higher performance than the
exclusive method on both metrics. In particular, Hybrid(S2
), which assigns a part of the CS to a
slice with low content popularity slope and low content request ratio, can most efficiently balance
both metrics when the part of the CS allocated to a slice is approximately one third (in general,
1/N) of the whole CS. From this result, we anticipate that even if the method used to
20
25
30
35
40
0 0.2 0.4 0.6 0.8 1
Average
content
delivery
time
[ms]
1
Hybrid(S1
)
Hybrid(S
2
)
Hybrid(S
3
)
Exclusive
Shared
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
Fairness
index
1
Hybrid(S1
)
Hybrid(S
2
)
Hybrid(S
3
)
Exclusive
Shared
Figure 10: Average content delivery time and fairness index against content request ratio r of VCCN slice
S1
for each allocation method when there is a difference of Zipf exponent α between VCCN slices.

14
20
25
30
35
40
0 500 1000 1500 2000 2500 3000
Avg.
content
delivery
time
[ms]
Size of CS allocated to R
n
on Hybrid(S
n
)
Hybrid(S1
)
Hybrid(S
2
)
Hybrid(S
3
)
Exclusive
Shared
0
0.2
0.4
0.6
0.8
1
0 500 1000 1500 2000 2500 3000
Fairness
index
Size of CS allocated to R
n
on Hybrid(S
n
)
Hybrid(S1
)
Hybrid(S
2
)
Hybrid(S
3
)
Exclusive
Shared
Figure 11: Average content delivery time and fairness index against the size of CS allocated to Rn
by
Hybrid(Sn
).
allocate the CS to VCCN slices is very simple, a hybrid method will be effective in terms of both
fairness and overall network performance.
Generally speaking, content request patterns in VCCN slices (i.e., the content popularity slope
and content request ratio for each VCCN slice) will be different. From the above results, the
method for allocating CS resources to VCCN routers should be selected as follows (see Table 1).
When the difference between content request patterns in VCCN slices is large, a hybrid method
that assigns a part of the CS to a slice with low content popularity slope and low content request
ratio will provide a balance between the two metrics. On the other hand, whenever fairness
between VCCN slices is not important for any content request patterns in the VCCN slices, the
shared method will maximize the performance of the entire network.
Table 1: The CS allocation method that should be selected under arbitrary content request patterns in terms
of trade-offs among the network fairness for VCCN slices and overall network performance
6. CONCLUSIONS
In this paper, we have analytically and quantitatively investigated the effects of CS allocation
methods and content request patterns in VCCN slices on network fairness for VCCN slices and
the overall network performance. We developed a mathematical model of virtualized CCN router
for cache performance analysis under arbitrary content request patterns, and derived the cache hit
rate for each VCCN router instance and the aggregated cache hit rate of the virtualized CCN
router. Furthermore, using several numerical examples, we have shown that when content request
patterns are heterogeneous, a hybrid resource allocation method will provide the best balance
between fairness and overall network performance.

15
In the future we will extend our model to investigate the effect of network topology in a network
composed of multiple virtualized CCN routers. We will also develop a dynamic CS allocation
method, which operates efficiently in a distributed environment in which both the content
popularity slope and the content request ratio of each VCCN slice are dynamically varying.
ACKNOWLEDGEMENTS
This research was partly supported by a Grant-in-Aid for Scientific Research (A) (24240010) and
(B) (25280030).
REFERENCE
[1] V. Jacobson, D. K. Smetters, J. D. Thornton, M. F. Plass, N. H. Briggs, and R. L. Braynard,
“Networking named content,” in Proc. the fifth International Conference on emerging Networking
EXperiments and Technologies (CoNEXT ’09), Dec. 2009, pp. 1–12.
[2] M. Ohtani, K. Tsukamoto, Y. Koizumi, H. Ohsaki, K. Hato, J. Murayama, and M. Imase, “VCCN:
virtual content-centric networking for realizing group-based communication,” in Proc. 12th IEEE
International Conference on Communications (ICC ’13), June 2013, pp. 2069–2073.
[3] G. Rossini and D. Rossi, “A dive into the caching performance of content centric networking,”
http://perso.telecom-paristech.fr/~drossi/paper/rossi11ccn-techrep2.pdf, Telecom ParisTech, Tech.
Rep.,
2011.
[4] D. Rossi and G. Rossini, “Caching performance of content centric networks under multi-path routing
(and more),” http://perso.telecom-paristech.fr/∼drossi/paper/rossi11ccn-techrep1.pdf, Telecom Paris-
Tech, Tech. Rep., 2011.
[5] I. Psaras, R. G. Clegg, R. Landa, W. K. Chai, and G. Pavlou, “Modelling and evaluation of CCN-
caching trees,” in Proc. the 10th international IFIP TC 6 conference on Networking - Volume Part I
(Networking ’11), May 2011, pp. 78–91.
[6] G. Carofiglio, V. Gehlen, and D. Perino, “Experimental evaluation of memory management in
content- centric networking,” in Proc. 10th IEEE International Conference on Communications (ICC
’11), June
2011, pp. 1–6.
[7] G. Carofiglio, M. Gallo, L. Muscariello, and D. Perino, “Evaluating per-application storage manage-
ment in content-centric networks,” Elsevier Computer Communications, vol. 36, no. 7, pp. 750–757,
Apr. 2013.
[8] C. Fricker, P. Robert, J. Roberts, and N. Sbihi, “Impact of traffic mix on caching performance in a
content-centric network,” in Proc. First Workshop on Emerging Design Choices in Name-Oriented
Networking at IEEE INFOCOM 2012 (NOMEN ’12), Mar. 2012, pp. 310–315.
[9] K. Ohsugi, K. Tsukamoto, and H. Ohsaki, “Study on the effect of CCN router virtualization on
content delivery time (in japanese),” IEICE Society Conference 2012, B-7-4, Aug. 2012.
[10] H. Che, Y. Tung, and Z. Wang, “Hierarchical Web caching systems: Modeling, design and exper-
imental results,” IEEE J. Selected Areas in Communications, vol. 20, no. 7, pp. 1305–1314, Sept.
2002.
[11] C. Fricker, P. Robert, and J. Roberts, “A versatile and accurate approximation for LRU cache perfor-
mance,” in Proc. 24th International Teletraffic Congress (ITC ’12), Sept. 2012, pp. 1–8.
[12] G. Carofiglio, M. Gallo, L. Muscariello, and D. Perino, “Modeling data transfer in content-centric
networking,” in Proc. the 23rd International Teletraffic Congress (ITC ’11), Sept. 2011, pp. 111–118.
[13] R. K. Jain, D.-M. W. Chiu, and W. R. Hawe, “A quantitative measure of fairness and discrimination
for resource allocation in shared computer systems,”
http://www1.cse.wustl.edu/∼jain/papers/ftp/fairness.pdf, Digital Equipment Corporation, Tech. Rep.
DEC-TR-301, 1984.

16
Authors
Keiichiro Tsukamoto received Bachelor of Information Science and Master of
Information Science degrees from Osaka University in 2009 and 2011,
respectively. He has been a graduate student of the Graduate School of Information
Science and Technology, Osaka University since April 2011. His research interests
are in the area of Web content mining and Content-Centric Networking. He is a
member of IEICE.
Kaito Ohsugi received Bachelor of Information Science from Osaka University in
2013. He has been a graduate student of the Graduate School of Information
Science and Technology, Osaka University since April 2013. He is a member of
IEICE.
Hiroyuki Ohsaki received the M. E. degree in the Information and Computer
Sciences from Osaka University, Osaka, Japan, in 1995. He also received the Ph.
D. degree from Osaka University, Osaka, Japan, in 1997. He is currently a
professor at Department of Informatics, School of Science and Technology,
Kwansei Gakuin University, Japan. His research work is in the area of design,
modeling, and control of large-scale communication networks. He is a member of
IEEE, IEICE, and IPSJ.
Toru Hasegawa received the B.E., the M.E. and Dr. Informatics degrees in
information engineering from Kyoto University, Japan, in 1982, 1984 and 2000,
respectively. Since joining KDD (currently KDDI) in 1984, he has been working in
the field of formal description technique (FDT) of communication protocols,
ATM and high-speed protocols. From 1990 to 1991, he was a visiting researcher
at Columbia University. His current interests are future Internets, Information
Centric Networking, mobile computing and so on. He is currently a professor an
Osaka university. He received the Meritorious Award on Radio of ARIB in 2003.
He is a fellow of IPSJ and IEICE.
Masayuki Murata received the M.E. and D.E. degrees in Information and
Computer Science from Osaka University, Japan, in 1984 and 1988, respectively.
In April 1984, he joined Tokyo Research Laboratory, IBM Japan, as a
Researcher. From September 1987 to January 1989, he was an Assistant
Professor with Computation Center, Osaka University. In February 1989, he
moved to the Department of Information and Computer Sciences, Faculty of
Engineering Science, Osaka University. In April 1999, he became a Professor of
Cybermedia Center, Osaka University, and is now with Graduate School of
Information Science and Technology, Osaka University since April 2004. His
research interests include computer communication networks, performance
modeling and evaluation. He is a member of IEEE, ACM and IEICE.

Cache Performance Analysis of Virtualized Router on Virtual Content-Centric Networks

Recommended

Recommended

More Related Content

Similar to Cache Performance Analysis of Virtualized Router on Virtual Content-Centric Networks

Similar to Cache Performance Analysis of Virtualized Router on Virtual Content-Centric Networks (20)

More from josephjonse

More from josephjonse (20)

Recently uploaded

Recently uploaded (20)

Cache Performance Analysis of Virtualized Router on Virtual Content-Centric Networks